mnm-sdk 0.1.1__tar.gz

mnm_sdk-0.1.1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Sohan Narewadi and Radhesh Nair
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

mnm_sdk-0.1.1/PKG-INFO

@@ -0,0 +1,73 @@
+Metadata-Version: 2.4
+Name: mnm-sdk
+Version: 0.1.1
+Summary: MNM client: log model predictions asynchronously, without slowing inference
+License: MIT
+Project-URL: Repository, https://github.com/Sohan-Narewadi/MNM
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Provides-Extra: dev
+Requires-Dist: pytest>=8; extra == "dev"
+Requires-Dist: ruff>=0.6; extra == "dev"
+Requires-Dist: black>=24; extra == "dev"
+Dynamic: license-file
+
+# sdk/ — MNM Python client
+
+A tiny library a model service imports to log predictions to MNM **without slowing down inference**.
+
+## Install
+```bash
+pip install mnm-sdk
+```
+
+## Usage
+```python
+from mnm_sdk import Monitor
+
+monitor = Monitor(api_url="http://backend:8000", model="churn", version="3")
+
+@app.post("/predict")
+def predict(req: Request):
+    t0 = time.perf_counter()
+    y = model.predict(req.features)
+    monitor.log(features=req.features, output=y, latency_ms=(time.perf_counter() - t0) * 1000)
+    return y
+```
+
+## How it behaves
+- `log()` only appends to an in-memory bounded queue and returns: it never waits on the network and never raises, whatever is passed in.
+- A background daemon thread sends batches to `POST /v1/ingest` when `batch_size` events are queued or every `flush_interval` seconds, whichever comes first.
+- A failed batch is retried with exponential backoff (`max_retries`, `retry_backoff`). Each event carries its own `request_id` and `ts`, and ingestion is idempotent on those, so a retry never stores an event twice. A batch the server rejects as invalid (4xx) is not retried.
+- `NaN` and `Infinity` (for example a missing feature from pandas or numpy) are sent as `null`, because JSON has no such numbers.
+- If the queue is full, the **oldest** event is dropped and counted. Nothing blocks.
+- `close()` runs at interpreter exit (`atexit`) and sends what is still queued.
+- `monitor.stats` reports `logged`, `sent`, `dropped`, `failed` and `queued`.
+- No dependencies outside the standard library. Python 3.9+.
+
+| Option | Default | Meaning |
+|---|---|---|
+| `batch_size` | 100 | Send as soon as this many events are queued |
+| `flush_interval` | 1.0 s | Send whatever is queued at least this often |
+| `max_queue` | 10 000 | Queue capacity; beyond it the oldest event is dropped |
+| `max_retries` | 3 | Extra attempts for a batch that failed |
+| `retry_backoff` | 0.5 s | Wait before the first retry; doubles each time |
+| `timeout` | 5.0 s | HTTP timeout per request |
+
+`log()` also accepts `request_id=`, `label=` and `ts=`. The event format is the ingest contract in [`docs/api-spec.md`](../docs/api-spec.md).
+
+## Development
+```bash
+cd sdk
+pip install -e ".[dev]"
+pytest          # no network and no backend needed
+```
+
+## Layout
+```
+sdk/
+├── mnm_sdk/__init__.py, monitor.py, transport.py
+├── tests/
+└── pyproject.toml
+```

mnm_sdk-0.1.1/README.md

@@ -0,0 +1,58 @@
+# sdk/ — MNM Python client
+
+A tiny library a model service imports to log predictions to MNM **without slowing down inference**.
+
+## Install
+```bash
+pip install mnm-sdk
+```
+
+## Usage
+```python
+from mnm_sdk import Monitor
+
+monitor = Monitor(api_url="http://backend:8000", model="churn", version="3")
+
+@app.post("/predict")
+def predict(req: Request):
+    t0 = time.perf_counter()
+    y = model.predict(req.features)
+    monitor.log(features=req.features, output=y, latency_ms=(time.perf_counter() - t0) * 1000)
+    return y
+```
+
+## How it behaves
+- `log()` only appends to an in-memory bounded queue and returns: it never waits on the network and never raises, whatever is passed in.
+- A background daemon thread sends batches to `POST /v1/ingest` when `batch_size` events are queued or every `flush_interval` seconds, whichever comes first.
+- A failed batch is retried with exponential backoff (`max_retries`, `retry_backoff`). Each event carries its own `request_id` and `ts`, and ingestion is idempotent on those, so a retry never stores an event twice. A batch the server rejects as invalid (4xx) is not retried.
+- `NaN` and `Infinity` (for example a missing feature from pandas or numpy) are sent as `null`, because JSON has no such numbers.
+- If the queue is full, the **oldest** event is dropped and counted. Nothing blocks.
+- `close()` runs at interpreter exit (`atexit`) and sends what is still queued.
+- `monitor.stats` reports `logged`, `sent`, `dropped`, `failed` and `queued`.
+- No dependencies outside the standard library. Python 3.9+.
+
+| Option | Default | Meaning |
+|---|---|---|
+| `batch_size` | 100 | Send as soon as this many events are queued |
+| `flush_interval` | 1.0 s | Send whatever is queued at least this often |
+| `max_queue` | 10 000 | Queue capacity; beyond it the oldest event is dropped |
+| `max_retries` | 3 | Extra attempts for a batch that failed |
+| `retry_backoff` | 0.5 s | Wait before the first retry; doubles each time |
+| `timeout` | 5.0 s | HTTP timeout per request |
+
+`log()` also accepts `request_id=`, `label=` and `ts=`. The event format is the ingest contract in [`docs/api-spec.md`](../docs/api-spec.md).
+
+## Development
+```bash
+cd sdk
+pip install -e ".[dev]"
+pytest          # no network and no backend needed
+```
+
+## Layout
+```
+sdk/
+├── mnm_sdk/__init__.py, monitor.py, transport.py
+├── tests/
+└── pyproject.toml
+```

mnm_sdk-0.1.1/mnm_sdk/__init__.py

@@ -0,0 +1,6 @@
+"""MNM SDK: log model predictions to MNM without slowing down inference."""
+
+from mnm_sdk.monitor import Monitor, Stats
+from mnm_sdk.transport import HttpTransport, Transport, TransportError
+
+__all__ = ["HttpTransport", "Monitor", "Stats", "Transport", "TransportError"]

mnm_sdk-0.1.1/mnm_sdk/monitor.py

@@ -0,0 +1,212 @@
+"""Monitor: the object a model service uses to log predictions.
+
+`log()` only appends to a bounded in-memory queue. A background thread sends
+the queue to the MNM ingestion API in batches. Nothing in here may slow down
+or break the caller's prediction path: `log()` never blocks on the network
+and never raises.
+"""
+
+from __future__ import annotations
+
+import atexit
+import logging
+import threading
+import uuid
+from collections import deque
+from collections.abc import Mapping
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from typing import Any
+
+from mnm_sdk.transport import HttpTransport, Transport, TransportError
+
+logger = logging.getLogger("mnm")
+
+
+@dataclass(frozen=True)
+class Stats:
+    """Counters since the Monitor was created."""
+
+    logged: int  # events accepted by log()
+    sent: int  # events delivered to the ingestion API
+    dropped: int  # events discarded before sending (queue full, invalid, closed)
+    failed: int  # events given up on after the retries ran out
+    queued: int  # events waiting in the queue right now
+
+
+class Monitor:
+    """Logs predictions of one model version to MNM in the background.
+
+    Args:
+        api_url: Base URL of the MNM backend, e.g. ``http://backend:8000``.
+        model: Model name.
+        version: Model version.
+        batch_size: Send as soon as this many events are queued.
+        flush_interval: Send whatever is queued at least this often (seconds).
+        max_queue: Queue capacity. When full, the oldest event is dropped.
+        max_retries: Extra attempts for a batch that failed to send.
+        retry_backoff: Wait before the first retry (seconds); doubles each time.
+        timeout: HTTP timeout per request (seconds).
+        transport: Replace the HTTP transport (used by tests).
+    """
+
+    def __init__(
+        self,
+        api_url: str,
+        model: str,
+        version: str | int,
+        *,
+        batch_size: int = 100,
+        flush_interval: float = 1.0,
+        max_queue: int = 10_000,
+        max_retries: int = 3,
+        retry_backoff: float = 0.5,
+        timeout: float = 5.0,
+        transport: Transport | None = None,
+    ) -> None:
+        self._model = model
+        self._version = str(version)
+        self._batch_size = max(1, batch_size)
+        self._flush_interval = flush_interval
+        self._max_queue = max(1, max_queue)
+        self._max_retries = max(0, max_retries)
+        self._retry_backoff = retry_backoff
+        self._transport = transport or HttpTransport(api_url, timeout=timeout)
+
+        self._queue: deque[dict[str, Any]] = deque()
+        self._lock = threading.Lock()
+        self._drained = threading.Condition(self._lock)
+        self._wake = threading.Event()
+        self._closing = threading.Event()
+        self._in_flight = 0
+        self._logged = self._sent = self._dropped = self._failed = 0
+
+        self._thread = threading.Thread(
+            target=self._run, name="mnm-monitor", daemon=True
+        )
+        self._thread.start()
+        atexit.register(self.close)
+
+    # -- called on the prediction path -------------------------------------
+
+    def log(
+        self,
+        features: Mapping[str, Any],
+        output: Any,
+        latency_ms: float,
+        *,
+        request_id: str | None = None,
+        label: Any = None,
+        ts: datetime | None = None,
+    ) -> None:
+        """Queue one prediction. Returns immediately and never raises."""
+        try:
+            event: dict[str, Any] = {
+                "model": self._model,
+                "version": self._version,
+                "request_id": request_id or uuid.uuid4().hex,
+                "ts": (ts or datetime.now(timezone.utc)).isoformat(),
+                "features": dict(features),
+                "output": output,
+                "latency_ms": float(latency_ms),
+            }
+            if label is not None:
+                event["label"] = label
+        except Exception:
+            with self._lock:
+                self._dropped += 1
+            logger.debug("mnm: dropped an event that could not be built", exc_info=True)
+            return
+
+        with self._lock:
+            if self._closing.is_set():
+                self._dropped += 1
+                return
+            if len(self._queue) >= self._max_queue:
+                self._queue.popleft()
+                self._dropped += 1
+            self._queue.append(event)
+            self._logged += 1
+            batch_ready = len(self._queue) >= self._batch_size
+        if batch_ready:
+            self._wake.set()
+
+    # -- control -------------------------------------------------------------
+
+    @property
+    def stats(self) -> Stats:
+        with self._lock:
+            return Stats(
+                logged=self._logged,
+                sent=self._sent,
+                dropped=self._dropped,
+                failed=self._failed,
+                queued=len(self._queue),
+            )
+
+    def flush(self, timeout: float | None = None) -> bool:
+        """Block until everything logged so far was sent or given up on.
+
+        Returns False if that did not happen within `timeout` seconds.
+        """
+        self._wake.set()
+        with self._drained:
+            return self._drained.wait_for(
+                lambda: not self._queue and not self._in_flight, timeout
+            )
+
+    def close(self, timeout: float = 5.0) -> None:
+        """Send what is queued, then stop the background thread. Idempotent."""
+        self._closing.set()
+        self._wake.set()
+        self._thread.join(timeout)
+
+    # -- background thread ---------------------------------------------------
+
+    def _run(self) -> None:
+        while True:
+            self._wake.wait(self._flush_interval)
+            self._wake.clear()
+            self._drain()
+            if self._closing.is_set():
+                self._drain()  # anything logged while the last batch was sending
+                return
+
+    def _drain(self) -> None:
+        while True:
+            with self._lock:
+                if not self._queue:
+                    self._drained.notify_all()
+                    return
+                size = min(self._batch_size, len(self._queue))
+                batch = [self._queue.popleft() for _ in range(size)]
+                self._in_flight = size
+            delivered = self._send(batch)
+            with self._lock:
+                self._in_flight = 0
+                if delivered:
+                    self._sent += len(batch)
+                else:
+                    self._failed += len(batch)
+                self._drained.notify_all()
+
+    def _send(self, batch: list[dict[str, Any]]) -> bool:
+        """Try to deliver one batch, retrying with backoff. Never raises."""
+        delay = self._retry_backoff
+        for attempt in range(self._max_retries + 1):
+            try:
+                self._transport.send(batch)
+                return True
+            except TransportError as error:
+                reason: Exception = error
+                if not error.retryable:
+                    break
+            # A bug in a transport must not kill the logging thread.
+            except Exception as error:  # noqa: BLE001
+                reason = error
+            if attempt < self._max_retries:
+                # Returns early when close() is called: do not delay shutdown.
+                self._closing.wait(delay)
+                delay *= 2
+        logger.warning("mnm: gave up on %d events: %s", len(batch), reason)
+        return False

mnm_sdk-0.1.1/mnm_sdk/transport.py

@@ -0,0 +1,76 @@
+"""How batches leave the process: HTTP POST to the MNM ingestion API."""
+
+from __future__ import annotations
+
+import json
+import urllib.error
+import urllib.request
+from typing import Any, Protocol
+
+INGEST_PATH = "/v1/ingest"
+
+
+class TransportError(Exception):
+    """A batch could not be delivered.
+
+    `retryable` is False when sending the same batch again cannot succeed
+    (the server rejected it as invalid).
+    """
+
+    def __init__(self, message: str, *, retryable: bool = True) -> None:
+        super().__init__(message)
+        self.retryable = retryable
+
+
+class Transport(Protocol):
+    def send(self, events: list[dict[str, Any]]) -> None:
+        """Deliver one batch or raise TransportError."""
+
+
+def _json_default(value: Any) -> Any:
+    """Encode values `json` does not know, e.g. numpy scalars and arrays."""
+    for method in ("item", "tolist"):
+        convert = getattr(value, method, None)
+        if callable(convert):
+            return convert()
+    return str(value)
+
+
+def encode(events: list[dict[str, Any]]) -> bytes:
+    """Serialise a batch as strict JSON.
+
+    NaN and +/-Infinity are not JSON. Python would write them as bare tokens,
+    which the ingestion database rejects for the whole batch, so they are sent
+    as null (a missing value). The round trip also catches the ones that only
+    appear after numpy values have been converted.
+    """
+    loose = json.dumps(events, default=_json_default)
+    if "NaN" not in loose and "Infinity" not in loose:
+        return loose.encode("utf-8")
+    strict = json.loads(loose, parse_constant=lambda _token: None)
+    return json.dumps(strict, allow_nan=False).encode("utf-8")
+
+
+class HttpTransport:
+    def __init__(self, api_url: str, *, timeout: float = 5.0) -> None:
+        self._url = api_url.rstrip("/") + INGEST_PATH
+        self._timeout = timeout
+
+    def send(self, events: list[dict[str, Any]]) -> None:
+        request = urllib.request.Request(
+            self._url,
+            data=encode(events),
+            headers={"Content-Type": "application/json"},
+            method="POST",
+        )
+        try:
+            with urllib.request.urlopen(request, timeout=self._timeout) as response:
+                response.read()
+        except urllib.error.HTTPError as error:
+            # 4xx means the batch itself is wrong, except "slow down" answers.
+            retryable = error.code >= 500 or error.code in (408, 429)
+            raise TransportError(
+                f"MNM ingest answered HTTP {error.code}", retryable=retryable
+            ) from error
+        except OSError as error:  # URLError, timeouts, refused connections
+            raise TransportError(f"MNM ingest unreachable: {error}") from error

mnm_sdk-0.1.1/mnm_sdk.egg-info/PKG-INFO

@@ -0,0 +1,73 @@
+Metadata-Version: 2.4
+Name: mnm-sdk
+Version: 0.1.1
+Summary: MNM client: log model predictions asynchronously, without slowing inference
+License: MIT
+Project-URL: Repository, https://github.com/Sohan-Narewadi/MNM
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Provides-Extra: dev
+Requires-Dist: pytest>=8; extra == "dev"
+Requires-Dist: ruff>=0.6; extra == "dev"
+Requires-Dist: black>=24; extra == "dev"
+Dynamic: license-file
+
+# sdk/ — MNM Python client
+
+A tiny library a model service imports to log predictions to MNM **without slowing down inference**.
+
+## Install
+```bash
+pip install mnm-sdk
+```
+
+## Usage
+```python
+from mnm_sdk import Monitor
+
+monitor = Monitor(api_url="http://backend:8000", model="churn", version="3")
+
+@app.post("/predict")
+def predict(req: Request):
+    t0 = time.perf_counter()
+    y = model.predict(req.features)
+    monitor.log(features=req.features, output=y, latency_ms=(time.perf_counter() - t0) * 1000)
+    return y
+```
+
+## How it behaves
+- `log()` only appends to an in-memory bounded queue and returns: it never waits on the network and never raises, whatever is passed in.
+- A background daemon thread sends batches to `POST /v1/ingest` when `batch_size` events are queued or every `flush_interval` seconds, whichever comes first.
+- A failed batch is retried with exponential backoff (`max_retries`, `retry_backoff`). Each event carries its own `request_id` and `ts`, and ingestion is idempotent on those, so a retry never stores an event twice. A batch the server rejects as invalid (4xx) is not retried.
+- `NaN` and `Infinity` (for example a missing feature from pandas or numpy) are sent as `null`, because JSON has no such numbers.
+- If the queue is full, the **oldest** event is dropped and counted. Nothing blocks.
+- `close()` runs at interpreter exit (`atexit`) and sends what is still queued.
+- `monitor.stats` reports `logged`, `sent`, `dropped`, `failed` and `queued`.
+- No dependencies outside the standard library. Python 3.9+.
+
+| Option | Default | Meaning |
+|---|---|---|
+| `batch_size` | 100 | Send as soon as this many events are queued |
+| `flush_interval` | 1.0 s | Send whatever is queued at least this often |
+| `max_queue` | 10 000 | Queue capacity; beyond it the oldest event is dropped |
+| `max_retries` | 3 | Extra attempts for a batch that failed |
+| `retry_backoff` | 0.5 s | Wait before the first retry; doubles each time |
+| `timeout` | 5.0 s | HTTP timeout per request |
+
+`log()` also accepts `request_id=`, `label=` and `ts=`. The event format is the ingest contract in [`docs/api-spec.md`](../docs/api-spec.md).
+
+## Development
+```bash
+cd sdk
+pip install -e ".[dev]"
+pytest          # no network and no backend needed
+```
+
+## Layout
+```
+sdk/
+├── mnm_sdk/__init__.py, monitor.py, transport.py
+├── tests/
+└── pyproject.toml
+```

mnm_sdk-0.1.1/mnm_sdk.egg-info/SOURCES.txt

@@ -0,0 +1,13 @@
+LICENSE
+README.md
+pyproject.toml
+mnm_sdk/__init__.py
+mnm_sdk/monitor.py
+mnm_sdk/transport.py
+mnm_sdk.egg-info/PKG-INFO
+mnm_sdk.egg-info/SOURCES.txt
+mnm_sdk.egg-info/dependency_links.txt
+mnm_sdk.egg-info/requires.txt
+mnm_sdk.egg-info/top_level.txt
+tests/test_monitor.py
+tests/test_transport.py

mnm_sdk-0.1.1/mnm_sdk.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

mnm_sdk-0.1.1/mnm_sdk.egg-info/requires.txt

@@ -0,0 +1,5 @@
+
+[dev]
+pytest>=8
+ruff>=0.6
+black>=24

mnm_sdk-0.1.1/mnm_sdk.egg-info/top_level.txt

@@ -0,0 +1 @@
+mnm_sdk

mnm_sdk-0.1.1/pyproject.toml

@@ -0,0 +1,24 @@
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "mnm-sdk"
+version = "0.1.1"
+description = "MNM client: log model predictions asynchronously, without slowing inference"
+requires-python = ">=3.9"
+readme = "README.md"
+license = { text = "MIT" }
+dependencies = []
+
+[project.urls]
+Repository = "https://github.com/Sohan-Narewadi/MNM"
+
+[project.optional-dependencies]
+dev = ["pytest>=8", "ruff>=0.6", "black>=24"]
+
+[tool.setuptools.packages.find]
+include = ["mnm_sdk*"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]

mnm_sdk-0.1.1/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

mnm_sdk-0.1.1/tests/test_monitor.py

@@ -0,0 +1,270 @@
+"""Monitor behaviour, tested through a fake transport (no network)."""
+
+from __future__ import annotations
+
+import threading
+import time
+from collections.abc import Iterator
+from datetime import datetime
+from typing import Any
+
+import pytest
+
+from mnm_sdk import Monitor, TransportError
+
+
+class FakeTransport:
+    """Records every batch. Can be told to fail or to block."""
+
+    def __init__(self) -> None:
+        self.batches: list[list[dict[str, Any]]] = []
+        self.attempts = 0
+        self.failures: list[Exception] = []
+        self.gate = threading.Event()
+        self.gate.set()
+
+    def send(self, events: list[dict[str, Any]]) -> None:
+        self.gate.wait(timeout=10)
+        self.attempts += 1
+        if self.failures:
+            raise self.failures.pop(0)
+        self.batches.append(events)
+
+    @property
+    def events(self) -> list[dict[str, Any]]:
+        return [event for batch in self.batches for event in batch]
+
+
+@pytest.fixture
+def transport() -> FakeTransport:
+    return FakeTransport()
+
+
+@pytest.fixture
+def make_monitor(transport: FakeTransport) -> Iterator[Any]:
+    monitors: list[Monitor] = []
+
+    def factory(**options: Any) -> Monitor:
+        options.setdefault("retry_backoff", 0.01)
+        monitor = Monitor(
+            "http://mnm.invalid",
+            model="churn",
+            version="3",
+            transport=transport,
+            **options,
+        )
+        monitors.append(monitor)
+        return monitor
+
+    yield factory
+    transport.gate.set()
+    for monitor in monitors:
+        monitor.close(timeout=2)
+
+
+def log_n(monitor: Monitor, n: int) -> None:
+    for i in range(n):
+        monitor.log(features={"i": i}, output=i, latency_ms=1.0)
+
+
+def test_logged_event_reaches_the_transport_in_the_ingest_format(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    monitor = make_monitor()
+
+    monitor.log(features={"age": 41, "plan": "pro"}, output=0.83, latency_ms=3.2)
+    assert monitor.flush(timeout=2)
+
+    (event,) = transport.events
+    assert event["model"] == "churn"
+    assert event["version"] == "3"
+    assert event["features"] == {"age": 41, "plan": "pro"}
+    assert event["output"] == 0.83
+    assert event["latency_ms"] == 3.2
+    assert len(event["request_id"]) >= 16
+    assert datetime.fromisoformat(event["ts"]).tzinfo is not None
+    assert "label" not in event
+
+
+def test_caller_supplied_request_id_and_label_are_kept(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    monitor = make_monitor()
+
+    monitor.log(features={}, output=1, latency_ms=0.5, request_id="abc", label=1)
+    monitor.flush(timeout=2)
+
+    assert transport.events[0]["request_id"] == "abc"
+    assert transport.events[0]["label"] == 1
+
+
+def test_events_are_sent_in_batches_of_batch_size(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    transport.gate.clear()
+    monitor = make_monitor(batch_size=10, flush_interval=60)
+
+    log_n(monitor, 25)
+    transport.gate.set()
+    monitor.flush(timeout=2)
+
+    assert [len(batch) for batch in transport.batches] == [10, 10, 5]
+    assert [event["output"] for event in transport.events] == list(range(25))
+
+
+def test_a_partial_batch_is_sent_after_the_flush_interval(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    monitor = make_monitor(batch_size=100, flush_interval=0.05)
+
+    log_n(monitor, 3)
+
+    deadline = time.monotonic() + 2
+    while not transport.events and time.monotonic() < deadline:
+        time.sleep(0.01)
+    assert len(transport.events) == 3
+
+
+def test_log_does_not_wait_for_a_stuck_backend(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    transport.gate.clear()  # every send now hangs
+    monitor = make_monitor(batch_size=10, flush_interval=0.01)
+
+    started = time.perf_counter()
+    log_n(monitor, 1000)
+    elapsed = time.perf_counter() - started
+
+    assert elapsed < 0.5
+    assert transport.batches == []
+
+
+def test_a_full_queue_drops_the_oldest_events_and_counts_them(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    transport.gate.clear()
+    monitor = make_monitor(batch_size=5, flush_interval=0.01, max_queue=10)
+    monitor.log(features={}, output="first", latency_ms=1.0)
+    deadline = time.monotonic() + 2
+    while monitor.stats.queued and time.monotonic() < deadline:
+        time.sleep(0.005)  # wait until the worker holds "first" and hangs
+
+    log_n(monitor, 15)
+    assert monitor.stats.dropped == 5
+    transport.gate.set()
+    monitor.flush(timeout=2)
+
+    outputs = [event["output"] for event in transport.events]
+    assert outputs == ["first", *range(5, 15)]
+    assert monitor.stats.sent == 11
+
+
+def test_a_failed_send_is_retried_with_the_same_events(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    transport.failures = [TransportError("down"), TransportError("still down")]
+    monitor = make_monitor(max_retries=3)
+
+    monitor.log(features={}, output=1, latency_ms=1.0, request_id="only")
+    monitor.flush(timeout=2)
+
+    assert transport.attempts == 3
+    assert [event["request_id"] for event in transport.events] == ["only"]
+    assert monitor.stats.sent == 1
+    assert monitor.stats.failed == 0
+
+
+def test_a_batch_is_given_up_after_max_retries_and_counted(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    transport.failures = [TransportError("down")] * 3
+    monitor = make_monitor(max_retries=2)
+
+    monitor.log(features={}, output="lost", latency_ms=1.0)
+    monitor.flush(timeout=2)
+    monitor.log(features={}, output="kept", latency_ms=1.0)
+    monitor.flush(timeout=2)
+
+    assert transport.attempts == 4  # 1 try + 2 retries, then the next batch
+    assert [event["output"] for event in transport.events] == ["kept"]
+    assert monitor.stats.failed == 1
+    assert monitor.stats.sent == 1
+
+
+def test_a_rejected_batch_is_not_retried(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    transport.failures = [TransportError("422 invalid", retryable=False)]
+    monitor = make_monitor(max_retries=5)
+
+    monitor.log(features={}, output=1, latency_ms=1.0)
+    monitor.flush(timeout=2)
+
+    assert transport.attempts == 1
+    assert monitor.stats.failed == 1
+
+
+def test_an_unexpected_transport_crash_does_not_kill_logging(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    transport.failures = [RuntimeError("bug in transport")]
+    monitor = make_monitor(max_retries=0)
+
+    monitor.log(features={}, output="lost", latency_ms=1.0)
+    monitor.flush(timeout=2)
+    monitor.log(features={}, output="kept", latency_ms=1.0)
+    monitor.flush(timeout=2)
+
+    assert [event["output"] for event in transport.events] == ["kept"]
+
+
+def test_log_never_raises_on_bad_arguments(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    monitor = make_monitor()
+
+    monitor.log(features=None, output=1, latency_ms=1.0)  # type: ignore[arg-type]
+    monitor.log(features=object(), output=1, latency_ms="x")  # type: ignore[arg-type]
+    monitor.log(features={"ok": 1}, output=1, latency_ms=1.0)
+    monitor.flush(timeout=2)
+
+    assert [event["features"] for event in transport.events] == [{"ok": 1}]
+    assert monitor.stats.dropped == 2
+
+
+def test_later_changes_to_the_features_dict_do_not_leak_into_the_event(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    transport.gate.clear()
+    monitor = make_monitor()
+    features = {"age": 41}
+
+    monitor.log(features=features, output=1, latency_ms=1.0)
+    features["age"] = 99
+    transport.gate.set()
+    monitor.flush(timeout=2)
+
+    assert transport.events[0]["features"] == {"age": 41}
+
+
+def test_close_sends_what_is_still_queued(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    monitor = make_monitor(batch_size=100, flush_interval=60)
+    log_n(monitor, 7)
+
+    monitor.close(timeout=2)
+
+    assert len(transport.events) == 7
+
+
+def test_logging_after_close_is_ignored_and_counted(
+    make_monitor: Any, transport: FakeTransport
+) -> None:
+    monitor = make_monitor()
+    monitor.close(timeout=2)
+
+    monitor.log(features={}, output=1, latency_ms=1.0)
+
+    assert transport.events == []
+    assert monitor.stats.dropped == 1

mnm_sdk-0.1.1/tests/test_transport.py

@@ -0,0 +1,129 @@
+"""HttpTransport against a throwaway HTTP server on localhost."""
+
+from __future__ import annotations
+
+import json
+import threading
+from collections.abc import Iterator
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from typing import Any
+
+import pytest
+
+from mnm_sdk import TransportError
+from mnm_sdk.transport import HttpTransport
+
+
+class Recorder(HTTPServer):
+    status = 202
+    requests: list[dict[str, Any]]
+
+
+class Handler(BaseHTTPRequestHandler):
+    server: Recorder
+
+    def do_POST(self) -> None:
+        body = self.rfile.read(int(self.headers["Content-Length"]))
+        self.server.requests.append(
+            {
+                "path": self.path,
+                "content_type": self.headers["Content-Type"],
+                "body": json.loads(body),
+                "raw": body.decode("utf-8"),
+            }
+        )
+        self.send_response(self.server.status)
+        self.end_headers()
+        self.wfile.write(b"{}")
+
+    def log_message(self, *args: Any) -> None:
+        pass
+
+
+@pytest.fixture
+def server() -> Iterator[Recorder]:
+    recorder = Recorder(("127.0.0.1", 0), Handler)
+    recorder.requests = []
+    thread = threading.Thread(target=recorder.serve_forever, daemon=True)
+    thread.start()
+    yield recorder
+    recorder.shutdown()
+    recorder.server_close()
+
+
+def url(server: Recorder) -> str:
+    return f"http://127.0.0.1:{server.server_address[1]}"
+
+
+def test_send_posts_the_events_as_a_json_array_to_v1_ingest(server: Recorder) -> None:
+    events = [{"model": "churn", "features": {"age": 41}}]
+
+    HttpTransport(url(server) + "/").send(events)
+
+    (request,) = server.requests
+    assert request["path"] == "/v1/ingest"
+    assert request["content_type"] == "application/json"
+    assert request["body"] == events
+
+
+def test_values_json_cannot_encode_are_converted_not_fatal(server: Recorder) -> None:
+    class NumpyLikeScalar:
+        def item(self) -> float:
+            return 1.5
+
+    HttpTransport(url(server)).send(
+        [{"output": NumpyLikeScalar(), "features": {"x": {1, 2}.__class__}}]
+    )
+
+    body = server.requests[0]["body"]
+    assert body[0]["output"] == 1.5
+    assert isinstance(body[0]["features"]["x"], str)
+
+
+def test_a_server_error_is_retryable(server: Recorder) -> None:
+    server.status = 503
+
+    with pytest.raises(TransportError) as raised:
+        HttpTransport(url(server)).send([{}])
+
+    assert raised.value.retryable
+
+
+def test_a_validation_error_is_not_retryable(server: Recorder) -> None:
+    server.status = 422
+
+    with pytest.raises(TransportError) as raised:
+        HttpTransport(url(server)).send([{}])
+
+    assert not raised.value.retryable
+
+
+def test_an_unreachable_server_is_retryable() -> None:
+    with pytest.raises(TransportError) as raised:
+        HttpTransport("http://127.0.0.1:9", timeout=0.5).send([{}])
+
+    assert raised.value.retryable
+
+
+def test_nan_and_infinity_are_sent_as_null_because_json_has_neither(
+    server: Recorder,
+) -> None:
+    class NumpyLikeNan:
+        def item(self) -> float:
+            return float("nan")
+
+    HttpTransport(url(server)).send(
+        [
+            {
+                "features": {"age": float("nan"), "x": float("inf"), "ok": 1.5},
+                "output": [NumpyLikeNan(), float("-inf")],
+            }
+        ]
+    )
+
+    request = server.requests[0]
+    assert "NaN" not in request["raw"]
+    assert "Infinity" not in request["raw"]
+    assert request["body"] == [
+        {"features": {"age": None, "x": None, "ok": 1.5}, "output": [None, None]}
+    ]